Online Appendix: Spatial Diffusion of Economic Shocks in Networks

Online Appendix: Spatial Diﬀusion of Economic Shocks in Networks

Ashani Amarasinghe∗ Roland Hodler† Paul A. Raschky‡ Yves Zenou§

September 2018

Not for Publication

∗Department of Economics, Monash University; email: [email protected]. †Department of Economics, University of St.Gallen; CEPR, CESifo, and OxCarre; email: [email protected]. ‡Department of Economics, Monash University; email: [email protected]. §Department of Economics, Monash University; CEPR and IZA; email: [email protected].

A.1 Contents

A Additional Data Description

B Correlation between Subnational GDP and Subnational Nighttime Lights in Africa

C Main Result (complete table)

D Robustness: Province-Year Fixed Eﬀects

E Robustness: Temporal and Spatial Lags

F Robustness: Spatial Clustering of Standard Error

G Robustness: Grid Cells as Unit of Observation

H Robustness: Excluding Top Mineral Producers

I Robustness: Fiscal Channel

J Robustness: Contiguity Network

K Robustness: Geodesic Network without Adjustment for Variability in Altitude

L Robustness: Alternative Ethnicity Network

M Robustness: Networks within and across Countries

N Top-Ten Key Player Rankings for Populous Countries

O Top-Ten Rankings from Policy Experiments for Kenya and Nigeria

P Top-Ten Rankings from Policy Experiments for Other Populous Countries

A.2 A Additional Data Description

A.1 Subnational Districts and Countries

Our analysis focuses on 5944 subnational districts, i.e., ADM2 regions, in 53 countries across the African continent. The list of countries and the number of subnational regions belonging to each country appear in Table A1. The geographic dispersion of subnational regions is graphically represented in Figure A1.

Figure A1: Districts in Africa

A.3 Table A1: List of Countries

Country No. of districts 1 Algeria 1,504 2 Angola 163 3 Benin 76 4 Botswana 25 5 Burkina Faso 301 6 Burundi 133 7 Cameroon 58 8 Cape Verde 16 9 Central African Republic 51 10 Chad 53 11 Comoros 3 12 Ivory Coast 50 13 Democratic Republic of the Congo 38 14 Djibouti 11 15 Egypt 26 16 Equatorial Guinea 6 17 Eritrea 50 18 Ethiopia 72 19 Gabon 37 20 Gambia 13 21 Ghana 137 22 Guinea 34 23 Guinea-Bissau 37 24 Kenya 48 25 Lesotho 10 26 Liberia 66 27 Libya 32 28 Madagascar 22 29 Malawi 253 30 Mali 51 31 Mauritania 44 32 Mauritius 10 33 Morocco 54 34 Mozambique 128 35 Namibia 107 36 Niger 36 37 Nigeria 775 38 Republic of Congo 46 39 Rwanda 142 40 Sao Tome and Principe 2 41 Senegal 30 42 Sierra Leone 14 43 Somalia 74 44 South Africa 354 45 Sudan 26 46 Swaziland 4 47 Tanzania 136 48 Togo 21 49 Tunisia 267 50 Uganda 162 51 Western Sahara 4 52 Zambia 72 53 Zimbabwe 60

A.4 A.2 Ethnic Homelands

The ethnic connectivity matrix is based on the digitized map version of the Murdock (1958) map of the boundaries of ethnic homelands in Africa shown in Figure A2. Using the spatial overlay tool in ArcMap 10.2, we combined the ADM2 polylines from Figure A1 with the ethnic homeland polylines from Figure A2 to assign each district the ethnicity of the ethnic homeland that covers the largest area of this district.

Figure A2: Ethnic Homelands in Africa

A.5 A.3 Road Network

Figure A3 shows the network of primary and secondary roads (in purple) from the Open- StreetMap data, with the district boundaries in the background (in light-gray).

Figure A3: Primary and Secondary Roads in Africa

A.6 A.4 Mines & Minerals

Figure A4 shows the location of mines from the SNL Minings & Metals database, with the district boundaries in the background. Table A2 lists the diﬀerent types of minerals covered in this database as well as the source for the information on the world market price of these minerals.

Figure A4: Distribution of Mines in Africa

A.7 Table A2: List of Minerals

Name Measure Source 1 Antimony Tonnes USGS Commodity Prices 2 Bauxite Tonnes USGS Commodity Prices 3 Chromite Tonnes USGS Commodity Prices 4 Coal Tonnes World Bank 5 Cobalt Tonnes USGS Commodity Prices 6 Copper Tonnes SNL-Thomas Reuters 7 Diamond Carats USGS Commodity Prices 8 Gold Ounces SNL-Thomas Reuters 9 Graphite Tonnes USGS Commodity Prices 10 Ilmenite Tonnes USGS Commodity Prices 11 Iron Tonnes World Bank 12 Lanthanide Tonnes USGS Commodity Prices 13 Lead Tonnes World Bank 14 Lithium Tonnes USGS Commodity Prices 15 Managanese Tonnes USGS Commodity Prices 16 Nickel Tonnes World Bank 17 Niobium Tonnes USGS Commodity Prices 18 Palladium Ounces SNL-Thomas Reuters 19 Phosphate Tonnes World Bank 20 Platinum Ounces SNL-Thomas Reuters 21 Potash Tonnes World Bank 22 Rutile Tonnes USGS Commodity Prices 23 Silver Ounces World Bank 24 Tin Tonnes SNL-Thomas Reuters 25 Tantalum Tonnes USGS Commodity Prices 26 Tungsten Tonnes USGS Commodity Prices 27 Uranium Oxide Pounds International Monetary Fund 28 Vanadium Tonnes USGS Commodity Prices 29 Yttrium Tonnes SNL-Thomas Reuters 30 Zinc Tonnes SNL-Thomas Reuters 31 Zircon Tonnes USGS Commodity Prices

A.8 B Correlation between Subnational GDP and Sub- national Nighttime Lights in Africa

Hodler and Raschky (2014, Appendix B) document a strong correlation between GDP per capita and nighttime lights in subnational administrative regions using the subnational GDP data by Gennaioli et al. (2014). In Table B1, we replicate their analysis using their data, but restricting the sample to the 82 subnational regions from the nine African countries for which Gennaioli et al. (2014) provide subnational GDP data. These countries are: Benin, Egypt, Kenya, Lesotho, Morocco, Mozambique, Nigeria, South Africa, and Tanzania. Comparing the results reported in Table B1 with those in Hodler and Raschky (2014) suggests that the relation between subnational GDP per capita and subnational nighttime lights is very similar in Africa as elsewhere.

Table B1: Subnational GDP and Nighttime Lights in Africa

(1) (2) Lightict 0.291*** 0.354*** (0.005) (0.047) R-squared 0.688 0.688 Observations 1,200 1,200 Region FE NO YES Notes: Dependent variable is the logarithm of regional GDP per capita. OLS regressions. Lightict is the logarithm of average nighttime lights. Robust standard errors in parentheses. ***, **, * indicate sig- niﬁcance at the 1, 5 and 10%-level, respectively.

A.9 C Main Result (complete table)

Table C1: Connectivity based on ethnicity, geography and roads

(1) (2) (3) (4) (5) (6) (7) (8) OLS IV OLS IV OLS IV OLS IV Dependent variable: Lightict

Ethnicity W Lightjct 0.552*** 0.271 0.160*** 0.342*** (0.015) (0.176) (0.013) (0.122) Inv Dist W Lightjct 0.550*** 0.639*** 0.246*** 0.305** (0.012) (0.131) (0.011) (0.124) Inv Road W Lightjct 0.556*** 0.280** 0.393*** 0.361*** (0.010) (0.113) (0.015) (0.116) P opulationict 0.243*** 0.179*** 0.178*** -0.023 0.162*** -0.098*** 0.095*** -0.220*** (0.033) (0.035) (0.027) (0.031) (0.026) (0.033) (0.026) (0.036) Conflictict -0.011* -0.012** -0.011* -0.011* -0.010** -0.013** -0.008 -0.006 (0.006) (0.006) (0.006) (0.006) (0.005) (0.006) (0.005) (0.006) Ethnicity W P opulationjct -0.377*** 0.096 -0.171*** -0.256*** (0.041) (0.060) (0.035) (0.057) Ethnicity W Conflictjct -0.031*** -0.043** -0.019* -0.020 (0.011) (0.018) (0.011) (0.017) Inv Dist W P opulationjct -0.247*** 0.363*** -0.021 0.240*** (0.039) (0.042) (0.036) (0.046) Inv Dist W Conflictjct -0.018* -0.018 -0.009 -0.012 (0.011) (0.015) (0.011) (0.015) Inv Road W P opulationjct -0.128*** 0.582*** 0.048 0.475*** (0.035) (0.043) (0.041) (0.050) Inv Road W Conflictjct -0.004 -0.018 0.004 0.001 (0.009) (0.012) (0.010) (0.013) MPict 0.125*** 0.118*** 0.120*** 0.116*** 0.111*** 0.112*** 0.115*** 0.107*** (0.014) (0.015) (0.013) (0.014) (0.013) (0.014) (0.014) (0.016)

First stage: Dependent variable: Lightjct

MPjct 0.121*** 0.124*** 0.119*** 0.123*** (0.015) (0.014) (0.014) (0.015)

First-stage F-stat 63.83 85.00 71.90 63.64 Observations 101,048 101,048 101,048 101,048 101,048 101,048 101,048 101,048 District FE YES YES YES YES YES YES YES YES Country-year FE YES YES YES YES YES YES YES YES Additional controls YES YES YES YES YES YES YES YES Notes: This table corresponds to Table 2 in Section 6, but reports the coefficient estimates on all (second-stage) control variables. Even columns report standard fixed effects regressions with district and country-year fixed effects, and odd columns IV estimates. See Section 4 for the definitions of all variables. The first stage further includes the control variables indicated in Section 5. Standard errors, clustered at the network level, are in parentheses. ***, **, * indicate significance at the 1, 5 and 10% level, respectively.

A.10 D Robustness: Province-Year Fixed Eﬀects

Table D1: Province-Year Fixed Eﬀects

(1) (2) (3) (4) (5) (6) (7) (8) OLS IV OLS IV OLS IV OLS IV Dependent variable: Lightict

Ethnicity W Lightjct 0.044** 0.301 -0.029 0.469*** (0.019) (0.187) (0.019) (0.151) Inv Dist W Lightjct 0.118*** 0.748*** 0.037** 0.473*** (0.015) (0.119) (0.014) (0.101) Inv Road W Lightjct 0.220*** 0.779*** 0.214*** 0.484*** (0.013) (0.108) (0.017) (0.119) First stage: Dependent variable: Lightjct

MPjct 0.132*** 0.131*** 0.133*** 0.136*** (0.015) (0.014) (0.015) (0.016)

First-stage F-stat 74.15 89.56 84.44 69.15

Observations 101,048 101,048 101,048 101,048 101,048 101,048 101,048 101,048 District FE YES YES YES YES YES YES YES YES Province-year FE YES YES YES YES YES YES YES YES Additional controls YES YES YES YES YES YES YES YES Notes: Even columns report standard fixed effects regressions with district and province(ADM1)-year fixed effects, and odd columns IV estimates. See Section 4 for the definitions of all variables. Ethnicity W Lightjct is weighted Lightjct with weights based on the row-normalized ethnicity matrix. Inv Dist W Lightjct (Inv Road W Lightjct) is weighted Lightjct with weights based on the row-normalized matrix of the inverse altitude-adjusted geodesic distances (inverse road distances) truncated at 70km. Additional control variables are population, conflict and MPict as well as weighted population and conflict in districts j 6= i. MPjct is an interaction term based on cross-sectional information on the location of mines and time-varying world prices of the commodities produced in these mines (see equation (11)). The first stage further includes the control variables indicated in equation (10). Standard errors, clustered at the network level, are in parentheses. ***, **, * indicate significance at the 1, 5 and 10%-level, respectively.

A.11 E Robustness: Temporal and Spatial Lags

Table E1: Temporal and Spatial Lags

(1) (2) (3) (4) (5) (6) (7) (8) OLS IV OLS IV OLS IV OLS IV Dependent variable: Lightict

Ethnicity W Lightjct−1 0.350*** 0.457** 0.114*** 0.356*** (0.019) (0.200) (0.017) (0.130) Inv Dist W Lightjct−1 0.334*** 0.695*** 0.147*** 0.366*** (0.014) (0.138) (0.013) (0.132) Inv Road W Lightjct−1 0.333*** 0.279** 0.229*** 0.317*** (0.011) (0.117) (0.014) (0.122)

First stage: Dependent variable: Lightjct−1

MPjct−1 0.111*** 0.115*** 0.111*** 0.114*** (0.015) (0.014) (0.015) (0.016)

First-stage F-stat 51.56 65.42 56.89 49.56 Observations 101,048 101,048 101,048 101,048 101,048 101,048 101,048 101,048 District FE YES YES YES YES YES YES YES YES Country-year FE YES YES YES YES YES YES YES YES Additional controls YES YES YES YES YES YES YES YES Notes: Even columns report standard fixed effects regressions with district and country-year fixed effects, and odd columns IV estimates. See Section 4 for the definitions of all variables. Ethnicity W Lightjct is weighted Lightjct with weights based on the row-normalized ethnicity matrix. Inv Dist W Lightjct (Inv Road W Lightjct) is weighted Lightjct with weights based on the row-normalized matrix of the inverse altitude-adjusted geodesic distances (inverse road distances) truncated at 70km. Additional control variables are population, conflict and MPict as well as weighted population and conflict in districts j 6= i. MPjct is an interaction term based on cross-sectional information on the location of mines and time-varying world prices of the commodities produced in these mines (see equation (11)). The first stage further includes the control variables indicated in equation (10). Standard errors, clustered at the network level, are in parentheses. ***, **, * indicate significance at the 1, 5 and 10%-level, respectively.

A.12 F Robustness: Spatial Clustering of Standard Errors

Table F1: Spatial Clustering of Standard Errors

(1) (2) (3) (4) (5) (6) (7) (8) OLS IV OLS IV OLS IV OLS IV Dependent variable: Lightict

Ethnicity W Lightjct 0.552*** 0.271 0.160*** 0.342*** (0.015) (0.165) (0.012) (0.107) Inv Dist W Lightjct 0.550*** 0.639*** 0.246*** 0.305*** (0.012) (0.132) (0.011) (0.112) Inv Road W Lightjct 0.556*** 0.280** 0.393*** 0.361*** (0.012) (0.126) (0.013) (0.108) First stage: Dependent variable: Lightjct

MPjct 0.121*** 0.124*** 0.119*** 0.123*** (0.013) (0.014) (0.013) (0.013)

First-stage F-stat 82.84 84.89 79.70 84.66

Observations 101,048 101,048 101,048 101,048 101,048 101,048 101,048 101,048 District FE YES YES YES YES YES YES YES YES Country-year FE YES YES YES YES YES YES YES YES Additional controls YES YES YES YES YES YES YES YES Notes: Even columns report standard fixed effects regressions with district and country-year fixed effects, and odd columns report IV estimates. See Section 4 for the definitions of all variables. Ethnicity W Lightjct is weighted Lightjct, with weights based on the row-normalized ethnicity matrix. Inv Dist W Lightjct (Inv Road W Lightjct) is weighted Lightjct, with weights based on the row-normalized matrix of the inverse altitude-adjusted geodesic distances (inverse road distances) truncated at 70km. Additional control variables are population, conflict and MPict as well as weighted population and conflict in districts j 6= i. MPjct is an interaction term based on cross-sectional information on the location of mines and time-varying world prices of the commodities produced in these mines (see equation (11)). The first stage further includes the control variables indicated in equation (10). Spatially clustered standard errors are in parentheses, allowing for spatial correlation up to 70km and for infinite serial correlation. ***, **, * indicate significance at the 1, 5 and 10% level, respectively.

A.13 G Robustness: Grid-Cells as Unit of Observation

Table G1: Grid-Cells as Unit of Observation

(1) (2) (3) (4) (5) (6) (7) (8) OLS IV OLS IV OLS IV OLS IV

Dependent variable: Lightict

Eth W Lightjct 0.517*** 0.356*** 0.168*** 0.392*** (0.010) (0.094) (0.010) (0.055) Inv Dist W Lightjct 0.518*** 0.422*** 0.338*** 0.250*** (0.007) (0.074) (0.012) (0.037) Inv Road W Lightjct 0.420*** 0.604*** 0.173*** 0.413*** (0.007) (0.053) (0.010) (0.067)

First stage: Dependent variable: Lightjct

MPjct 0.258*** 0.247*** 0.231*** 0.230*** (0.020) (0.018) (0.018) (0.020)

First-stage F-stat 165.30 197.93 172.06 128.58

Observations 175,695 175,695 175,695 175,695 175,695 175,695 175,695 175,695 District FE YES YES YES YES YES YES YES YES Country-year FE YES YES YES YES YES YES YES YES Additional controls YES YES YES YES YES YES YES YES Notes: The units of observation are rectangular grid cells of 0.5 × 0.5 degrees (i.e., around 55 × 55km at the equator). Even columns report standard fixed effects regressions with district and country-year fixed effects, and odd columns report IV estimates. See Section 4 for the definitions of all variables. Ethnicity W Lightjct is weighted Lightjct, with weights based on the row-normalized ethnicity matrix. Inv Dist W Lightjct (Inv Road W Lightjct) is weighted Lightjct, with weights based on the row-normalized matrix of the inverse altitude-adjusted geodesic distances (inverse road distances) truncated at 70km. Additional control variables are population, conflict and MPict as well as weighted population and conflict in districts j 6= i. MPjct is an interaction term based on cross-sectional information on the location of mines and time-varying world prices of the commodities produced in these mines (see equation (11)). The first stage further includes the control variables indicated in equation (10). Standard errors, clustered at the network level, are in parentheses. ***, **, * indicate significance at the 1, 5 and 10%-level, respectively.

A.14 H Robustness: Excluding Top Mineral Producers

Table H1: Dropping Large Players

(1) (2) (3) (4) (5) (6) (7) (8) OLS IV OLS IV OLS IV OLS IV Dependent variable: Lightict

Ethnicity W Lightjct 0.546*** 0.394*** 0.120*** 0.477*** (0.015) (0.125) (0.014) (0.102) Inv Dist W Lightjct 0.630*** 0.398*** 0.362*** -0.001 (0.012) (0.128) (0.015) (0.107) Inv Road W Lightjct 0.552*** 0.420*** 0.317*** 0.497*** (0.011) (0.092) (0.014) (0.098)

First stage: Dependent variable: Lightjct

MPjct 0.211*** 0.214*** 0.208*** 0.211*** (0.019) (0.018) (0.018) (0.019)

First-stage F-stat 127.84 149.95 132.30 118.38 Observations 95,030 95,030 95,030 95,030 95,030 95,030 95,030 95,030 District FE YES YES YES YES YES YES YES YES Country-year FE YES YES YES YES YES YES YES YES Additional controls YES YES YES YES YES YES YES YES Notes: Sample is restricted to districts of countries that do not belong to the top ten producers for any mineral under consideration over the period 1997–2013 (see Table A2). Even columns report standard fixed effects regressions with district and country-year fixed effects, and odd columns IV estimates. See Section 4 for the definitions of all variables. Ethnicity W Lightjct is weighted Lightjct with weights based on the row-normalized ethnicity matrix. Inv Dist W Lightjct (Inv Road W Lightjct) is weighted Lightjct with weights based on the row-normalized matrix of the inverse altitude-adjusted geodesic distances (inverse road distances) truncated at 70km. Additional control variables are population, conflict and MPict as well as weighted population and conflict in districts j 6= i. MPjct is an interaction term based on cross-sectional information on the location of mines and time-varying world prices of the commodities produced in these mines (see equation (11)). The first stage further includes the control variables indicated in equation (10). Standard errors, clustered at the network level, are in parentheses. ***, **, * indicate significance at the 1, 5 and 10%-level, respectively.

A.15 I Robustness: Fiscal Channel

Table I1: Controlling for Connectivity based on ADM1 Networks

(1) (2) (3) (4) (5) (6) (7) (8) OLS IV OLS IV OLS IV OLS IV Lightict Ethnicity W Lightjct 0.315*** 0.274 0.126*** 0.295** (0.014) (0.169) (0.013) (0.123) Inv Dist W Lightjct 0.386*** 0.522*** 0.209*** 0.288** (0.013) (0.127) (0.012) (0.114) Inv Road W Lightjct 0.441*** 0.396*** 0.360*** 0.382*** (0.012) (0.115) (0.015) (0.108) ADM1 W Lightjct 0.434*** 0.256 0.329*** 0.834*** 0.270*** 0.236 0.142*** 0.292** (0.018) (0.266) (0.014) (0.152) (0.013) (0.148) (0.014) (0.130)

First stage: Lightjct MPjct 0.123*** 0.126*** 0.122*** 0.124*** (0.015) (0.015) (0.015) (0.016) First-stage F-stat 63.59 74.76 66.71 62.21 Observations 101,048 101,048 101,048 101,048 101,048 101,048 101,048 101,048 District FE YES YES YES YES YES YES YES YES Country Year FE YES YES YES YES YES YES YES YES Additional Controls YES YES YES YES YES YES YES YES Notes: Even columns report standard fixed effects regressions with district and country-year fixed effects, and odd columns IV estimates. See Section 4 for the definitions of all variables. Ethnicity W Lightjct is weighted Lightjct with weights based on the row-normalized ethnicity matrix. Inv Dist W Lightjct (Inv Road W Lightjct) is weighted Lightjct with weights based on the row-normalized matrix of the inverse altitude-adjusted geodesic distances (inverse road distances) truncated at 70km. ADM1 W Lightjct is weighted Lightjct with weights based on the row-normalized ADM1 matrix, which identifies whether districts belong to the same ADM1 unit. Additional control variables are population, conflict and MPict as well as weighted population and conflict in districts j 6= i. MPjct is an interaction term based on cross-sectional information on the location of mines and time-varying world prices of the commodities produced in these mines (see equation (11)). The first stage further includes the control variables indicated in equation (10). Standard errors, clustered at the network level, are in parentheses. ***, **, * indicate significance at the 1, 5 and 10%-level, respectively.

A.16 J Robustness: Contiguity Network

Table J1: Connectivity based on contiguity

(1) (2) OLS IV Dependent variable: Lightict

Contiguity W Lightjct 0.746*** 0.607*** (0.008) (0.158)

First stage: Dependent variable: Lightjct

MPjct 0.127*** (0.014)

First-stage F-Stat 83.17 Observations 101,048 101,048 District FE YES YES Country-year FE YES YES Additional controls YES YES Notes: Column (1) reports the standard fixed effects re- gression with district and country-year fixed effects, and column (2) the IV estimates. See Section 4 for the definitions of all variables. Contiguity W Lightjct is weighted Lightjct with weights based on the row-normalized contiguity matrix. Additional control variables are population, conflict and MPict as well as weighted population and conflict in districts j 6= i. MPjct is an interaction term based on cross-sectional information on the location of mines and time-varying world prices of the commodities produced in these mines (see equation (11)). The first stage further includes the control variables indicated in equation (10). Standard errors, clustered at the network level, are in parentheses. ***, **, * indicate significance at the 1, 5 and 10%-level, respectively.

A.17 K Robustness: Geodesic Network without Adjust- ment for Variability in Altitude

Table K1: Geodesic Network without Adjustment for Variability in Altitude

(1) (2) (3) (4) OLS IV OLS IV Dependent variable: Lightict

Ethnicity W Lightjct 0.110*** 0.396*** (0.013) (0.124) Inv Dist W Lightjct 0.664*** 0.750*** 0.374*** 0.439*** (0.012) (0.156) (0.013) (0.132) Inv Road W Lightjct 0.330*** 0.396*** (0.014) (0.113)

First stage: Dependent variable: Lightjct

MPjct 0.127*** 0.124*** (0.014) (0.015)

First-stage F-stat 88.61 64.99 Observations 101,048 101,048 101,048 101,048 District FE YES YES YES YES Country-year FE YES YES YES YES Additional controls YES YES YES YES Notes: Even columns report standard fixed effects regressions with district and country-year fixed effects, and odd columns IV estimates. See Section 4 for the definitions of all variables. Ethnicity W Lightjct is weighted Lightjct with weights based on the row-normalized ethnicity matrix. Inv Dist W Lightjct (Inv Road W Lightjct) is weighted Lightjct with weights based on the row-normalized matrix of the inverse geodesic distances without adjustment for the variability in altitude (inverse road distances) truncated at 70km. Additional control variables are population, conflict and MPict as well as weighted population and conflict in districts j 6= i. MPjct is an interaction term based on cross-sectional information on the location of mines and time-varying world prices of the commodities produced in these mines (see equation (11)). The first stage further includes the control variables indicated in equation (10). Standard errors, clustered at the network level, are in parentheses. ***, **, * indicate significance at the 1, 5 and 10%-level, respectively.

A.18 L Robustness: Alternative Ethnicity Network

Table L1: Alternative Ethnicity Network

(1) (2) (3) (4) OLS IV OLS IV Dependent variable: Lightict

Ethnicity W Lightjct 0.750*** 1.146 0.189*** 0.717 (0.032) (0.710) (0.017) (0.467) Inv Dist W Lightjct 0.283*** 0.271** (0.013) (0.109) Inv Road W Lightjct 0.410*** 0.426*** (0.017) (0.112) First stage: Dependent variable: Lightjct

MPjct 0.121*** 0.123*** (0.024) (0.023)

First-stage F-stat 26.72 29.07

Observations 101,048 101,048 101,048 101,048 District FE YES YES YES YES Country-year FE YES YES YES YES Additional controls YES YES YES YES Notes: Even columns report standard fixed effects regressions with district and country-year fixed effects, and odd columns IV estimates. See Section 4 for the definitions of all variables. Ethnicity W Lightjct is weighted Lightjct with weights based on the row-normalized alternative ethnicity matrix, as discussed in Section 6.2. Inv Dist W Lightjct (Inv Road W Lightjct) is weighted Lightjct with weights based on the row-normalized matrix of the inverse altitude-adjusted geodesic distances (inverse road distances) truncated at 70km. Additional control variables are population, conflict and MPict as well as weighted population and conflict in districts j 6= i. MPjct is an interaction term based on cross- sectional information on the location of mines and time-varying world prices of the commodities produced in these mines (see equation (13)). The first stage further includes the control variables indicated in equation (12). Standard errors, clustered at the network level, are in parentheses. ***, **, * indicate significance at the 1, 5 and 10%-level, respectively.

A.19 M Networks within and across Countries

We create two new sets of connectivity matrices: The first one only includes connected districts j which are in the same country as district i (Within Country), while the second one only includes connected districts j that are in other countries than district i (Outside Country). Table M1 below looks at only within country connectivity. Accordingly, the ethnicity matrix here only captures districts which are of the same ethnicity and belong to the same country. The inverse distance (road) matrix captures districts where the geodesic (road) distance is less than 70km and which belong to the same country. Compared to the main specification that allows for spill-overs within and between countries, the estimated ρ’s for ethnic and inverse distance connectivity are also positive but smaller in magnitude and no longer statically significant. In contrast, the ρ for road connectivity is larger in magnitude and highly statistically significant. Table M2 isolates outside-country spillover effects. Accordingly, the ethnicity matrix here only captures district which are of the same ethnicity and which do not belong to the same country. The inverse distance (road) matrix captures districts where the geodesic (road) distance is less than 70km, and do not belong to the same country. Here a different pattern emerges. Ethnic and geographic connectivity are more important for between country spill-overs, while the effect of road-connectivity is smaller and not statistically significant.

A.20 Table M1: Within-Country Networks

(1) (2) (3) (4) (5) (6) (7) (8) OLS IV OLS IV OLS IV OLS IV Lightict Ethnicity W Lightjct 0.484*** -0.035 0.144*** 0.147 (W ithin Country) (0.014) (0.140) (0.013) (0.101)

Inv Dist W Lightjct 0.504*** 0.393*** 0.236*** 0.115 (W ithin Country) (0.011) (0.110) (0.012) (0.113)

Inv Road W Lightjct 0.479*** 0.278*** 0.318*** 0.482*** (W ithin Country) (0.011) (0.103) (0.015) (0.106)

First stage: Lightjct MPjct 0.121*** 0.128*** 0.117*** 0.122*** (0.014) (0.014) (0.014) (0.015)

First-stage F-stat 75.79 90.17 70.16 68.83 Observations 101,048 101,048 101,048 101,048 101,048 101,048 101,048 101,048 District FE YES YES YES YES YES YES YES YES Country-year FE YES YES YES YES YES YES YES YES Additional controls YES YES YES YES YES YES YES YES Notes: Even columns report standard fixed effects regressions with district and country-year fixed effects, and odd columns IV estimates. See Section 4 for the definitions of all variables. Ethnicity W Lightjct (W ithin Country) is weighted Lightjct for districts belonging to the same country, with weights based on the row-normalized ethnicity matrix. Inv Dist W Lightjct (W ithin Country)(Inv Road W Lightjct (W ithin Country)) is weighted Lightjct for districts belonging to the same country, with weights based on the row-normalized matrix of the inverse altitude-adjusted geodesic distances (inverse road distances) truncated at 70km. Additional control variables are population, conflict and MPict as well as weighted population and conflict in districts j 6= i. MPjct is an interaction term based on cross-sectional information on the location of mines and time-varying world prices of the commodities produced in these mines (see equation (13)). The first stage further includes the control variables indicated in equation (12). Standard errors, clustered at the network level, are in parentheses. ***, **, * indicate significance at the 1, 5 and 10%-level, respectively.

A.21 Table M2: Outside-Country Networks

(1) (2) (3) (4) (5) (6) (7) (8) OLS IV OLS IV OLS IV OLS IV Lightict Ethnicity W Lightjct 0.262*** 0.291** 0.177*** 0.382*** (Outside Country) (0.010) (0.146) (0.010) (0.096)

Inv Dist W Lightjct 0.221*** 0.504*** 0.102*** 0.678*** (Outside Country) (0.010) (0.140) (0.011) (0.123)

Inv Road W Lightjct 0.228*** 0.380*** 0.151*** 0.144 (Outside Country) (0.009) (0.111) (0.010) (0.105)

First stage: Lightjct MPjct 0.127*** 0.126*** 0.116*** 0.126*** (0.013) (0.012) (0.013) (0.013)

First-stage F-stat 91.73 105.79 86.46 90.49

Observations 101,048 101,048 101,048 101,048 101,048 101,048 101,048 101,048 District FE YES YES YES YES YES YES YES YES Country-year FE YES YES YES YES YES YES YES YES Additional controls YES YES YES YES YES YES YES YES Notes: Even columns report standard fixed effects regressions with district and country-year fixed effects, and odd columns IV estimates. See Section 4 for the definitions of all variables. Ethnicity W Lightjct (Outside Country) is weighted Lightjct for districts not belonging to the same country, with weights based on the row-normalized ethnicity matrix. Inv Dist W Lightjct (Outside Country)(Inv Road W Lightjct (Outside Country)) is weighted Lightjct for districts not belonging to the same country, with weights based on the row-normalized matrix of the inverse altitude-adjusted geodesic distances (inverse road distances) truncated at 70km. Additional control variables are population, conflict and MPict as well as weighted population and conflict in districts j 6= i. MPjct is an interaction term based on cross-sectional information on the location of mines and time-varying world prices of the commodities produced in these mines (see equation (13)). The first stage further includes the control variables indicated in equation (12). Standard errors, clustered at the network level, are in parentheses. ***, **, * indicate significance at the 1, 5 and 10%-level, respectively.

A.22 N Top-Ten Key Player Rankings for Populous Coun- tries

Table N1: Top-Ten Key Player Rankings for Populous Countries

(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) Country Province District Overall Overall Overall Overall Ethnicity Road Inv.Dist KP Katz-Bon Betw. Eig. KP KP KP Rank Rank Rank Rank Rank Rank Rank Rank Ethiopia Ethiopia Addis Ababa Zone 4* 1 6 47 14 2 2 1 Ethiopia Addis Ababa Zone 3* 2 7 38 16 6 4 3 Ethiopia Addis Ababa Zone 2* 3 8 43 16 3 3 5 Ethiopia Addis Ababa Zone 6* 4 10 59 16 10 11 7 Ethiopia Oromia North Shewa (K4) 5 12 48 67 27 8 8 Ethiopia Addis Ababa Zone 5* 6 9 44 14 4 6 2 Ethiopia Tigray Mekele 7 2 49 70 5 7 6 Ethiopia Addis Ababa Addis Ababa* 8 4 38 16 1 1 4 Ethiopia Amhara West Gojam 9 17 10 9 18 9 9 Ethiopia Amhara Bar Dar Sp. Zone 10 14 59 57 7 10 10 Egypt Egypt Al Qalyubiyah 1 9 17 3 1 1 1 Egypt Al Gharbiyah 2 5 11 1 3 4 3 Egypt Al Minuﬁyah 3 10 16 6 2 3 4 Egypt Ash Sharqiyah 4 6 6 2 6 5 5 Egypt Ad Daqahliyah 5 2 8 5 9 11 2 Egypt Dumyat 6 1 22 20 8 9 11 Egypt Al Buhayrah 7 11 7 7 13 10 7 Egypt Al Qahirah* 8 8 15 8 10 6 6 Egypt Bani Suwayf 9 14 13 10 15 16 12 Egypt Kafr ash Shaykh 10 4 20 14 12 13 15 Democratic Republic of the Congo (DRC) DRC Kivu Sud-Kivu 1 3 9 2 20 1 20 DRC Bas-Congo Matadi 2 15 30 1 2 2 1 DRC Kasa¨ı-Oriental Mbuji-Mayi 3 5 30 2 3 5 4 DRC Kasa¨ı-Oriental Tshilenge 4 7 29 2 17 37 4 DRC Katanga Lubumbashi 5 11 30 2 1 3 2 DRC Kivu Bukavu 6 1 28 37 6 8 6 DRC Bas-Congo Boma 7 2 26 38 4 7 3 DRC Kinshasa City Kinshasa* 8 18 18 2 5 6 7 DRC Kivu Nord-Kivu 9 9 8 2 19 9 19 DRC Bas-Congo Bas-Fleuve 10 4 25 36 18 10 8 South Africa (SA) SA Gauteng Johannesburg* 1 82 228 4 1 1 1 SA Western Cape Kuils River* 2 263 339 228 10 3 5 SA Gauteng Roodepoort* 3 86 266 10 4 11 3 SA Western Cape Mitchells Plain* 4 261 344 228 8 2 8 SA Gauteng Benoni* 5 73 257 12 7 10 6 SA KwaZulu-Natal Durban 6 13 129 199 9 4 7 SA Gauteng Boksburg* 7 72 108 3 2 8 2 SA Gauteng Kempton Park* 8 77 200 5 17 6 9 SA Gauteng Soweto* 9 85 245 6 3 7 4 SA Gauteng Soshanguve* 10 104 313 26 14 12 13

A.23 Tanzania Tanzania Dar-Es-Salaam Ilala 1 124 39 82 5 3 1 Tanzania Dar-Es-Salaam Kinondoni 2 126 37 7 4 1 2 Tanzania Arusha Arusha 3 40 110 83 3 2 4 Tanzania Dar-Es-Salaam Temeke 4 125 119 86 7 4 5 Tanzania Arusha Arumeru 5 39 99 28 24 7 18 Tanzania Kilimanjaro Moshi Urban 6 31 111 14 6 5 6 Tanzania Kilimanjaro Moshi Rural 7 36 26 17 21 15 30 Tanzania Kaskazini-Unguja Kaskazini ’B’ 8 131 103 4 25 11 9 Tanzania Kilimanjaro Rombo 9 33 81 17 63 34 132 Tanzania Mwanza Nyamagana 10 28 114 111 2 6 3 Notes: Overall key-player (KP) rank is based on the ρs estimated in column (8) of Table 2. Overall Katz-Bonacich rank is based on the ρs estimated in column (8) of Table 2 and a weighting vector of 1. Ethnicity KP rank is based on ρ1 estimated in column (2) of Table 2. Inverse distance KP rank is based on ρ2 estimated in column (4) of Table 2. Road KP rank is based on ∗ ρ3 estimated in column (6) Table 2. indicate districts that are (part of) capital cities. South Africa has three capital cities i.e. Pretoria (Gauteng), Bloemfontein (Free State) and Cape Town (Western Cape). The number of districts per country are: Ethiopia 72, Egypt 26 (ADM1 level), DRC 38, South Africa 354, and Tanzania 136.

A.24 O Top-Ten Rankings from Policy Experiments for Kenya and Nigeria

Table O1 presents the ten districts in Nigeria and Kenya where a counterfactual increase in economic activity (see Section 8.1) would have the largest overall impact. Table O2 presents the ten districts in Nigeria and Kenya where a counterfactual improvement of the road connectivity (see Section 8.2) would have the largest overall impact.

Table O1: Top-Ten Rankings from Policy Experiment 1 – Nighttime Lights

(1) (2) (3) (4) (5) Country Province District Overall Overall Rank Key-Player Rank Nigeria Nigeria Lagos Surulere 1 9 Nigeria Lagos Mainland 2 8 Nigeria Lagos Shomolu 3 16 Nigeria Lagos Oshodi/Isolo 4 18 Nigeria Bayelsa Nembe 5 771 Nigeria Lagos Amuwo Odofin 6 10 Nigeria Lagos Alimosho 7 14 Nigeria Lagos Kosofe 8 25 Nigeria Bayelsa Brass 9 769 Nigeria Delta Warri South-West 10 772 Kenya Kenya Coast Lamu 1 48 Kenya Central Machakos 2 47 Kenya Coast Kilifi 3 46 Kenya Coast Kwale 4 3 Kenya Central Kiambu 5 5 Kenya Central Murang’a 6 7 Kenya Eastern Machakos 7 6 Kenya Eastern Embu 8 19 Kenya Rift Valley Nakuru 9 4 Kenya Central Kirinyaga 10 10 Notes: Overall rank reflects the district’s overall impact from increasing its average nighttime light pixel value by 10 on average nighttime lights across African districts (see Section 8.1 for a more detailed explanation). This counterfactual exercise is based on the ρs estimated in column (8) of Table 3. Nigeria has 775 districts, and Kenya has 48 districts.

A.25 Table O2: Top-Ten Rankings from Policy Experiment 2 – Roads

(1) (2) (3) (4) (5) Country Province District Overall Overall Rank Key-Player Rank Nigeria Nigeria Rivers Bonny 1 763 Nigeria Rivers Okrika 1 21 Nigeria Rivers Khana 3 99 Nigeria Delta Burutu 4 75 Nigeria Delta Warri North 5 79 Nigeria Rivers Andoni/O 6 127 Nigeria Delta Ughelli South 7 45 Nigeria Rivers Abua/Odu 8 42 Nigeria Rivers Akukutor 9 766 Nigeria Bayelsa Yenegoa 10 31 Kenya Kenya Central Machakos 1 47 Kenya Eastern Wajir 2 31 Kenya Eastern Meru 3 43 Kenya Central Nyeri 3 8 Kenya Central Kirinyaga 5 10 Kenya Central Murang’a 6 7 Kenya Eastern Machakos 6 6 Kenya Rift Valley Narok 8 9 Kenya Coast Kwale 9 3 Kenya Rift Valley Bomet 10 24 Kenya Rift Valley Nakuru 10 4 Notes: Overall rank reﬂects the district’s overall impact from adding a road link to the contiguous district with the highest average nighttime light pixel value to which there exists no road link (see Section 8.2 for a more detailed explanation). This counterfactual exercise is based on the ρs estimated in column (8) of Table 3. Nigeria has 775 districts, and Kenya has 48 districts.

A.26 P Top-Ten Rankings from Policy Experiments for Other Populous Countries

Table P1 presents the ten districts in Ethiopia, Egypt, DRC, South Africa, and Tanzania where a counterfactual increase in economic activity (see Section 8.1) would have the largest overall impact. Table P2 presents the ten districts in the same country where a counterfactual improvement of the road connectivity (see Section 8.2) would have the largest overall impact.

Table P1: Top-Ten Rankings from Policy Experiment 1 (Nighttime Lights) for Populous Countries

(1) (2) (3) (4) (5) Country Province District Overall Overall Rank KP Rank Ethiopia Ethiopia Amhara West Gojam 1 9 Ethiopia Addis Ababa Addis Ababa 2 8 Ethiopia Addis Ababa Zone 5 3 6 Ethiopia Addis Ababa Zone 4 4 1 Ethiopia Afar Zone 5 5 70 Ethiopia Oromia East Shewa 6 69 Ethiopia Oromia West Shewa 7 13 Ethiopia Amhara North Shewa (K3) 8 68 Ethiopia Addis Ababa Zone 2 9 3 Ethiopia Addis Ababa Zone 3 10 2 Egypt Egypt Suhaj 1 18 Egypt Al Jizah 2 24 Egypt Asyut 3 19 Egypt Qina 4 20 Egypt Al Minya 5 22 Egypt Matruh 6 26 Egypt As Suways 7 23 Egypt Al Bahr al Ahmar 8 25 Egypt Ad Daqahliyah 9 5 Egypt Al Wadi al Jadid 10 13

A.27 Democratic Republic of the Congo (DRC) DRC Katanga Haut-Shaba 1 37 DRC Equateur´ Sud-Ubangi 2 38 DRC Kivu Sud-Kivu 3 1 DRC Katanga Lubumbashi 4 5 DRC Kivu Nord-Kivu 5 9 DRC Bas-Congo Boma 6 7 DRC Kinshasa City Kinshasa 7 8 DRC Bandundu Mai-Ndombe 8 32 DRC Kasa¨ı-Oriental Tshilenge 9 4 DRC Bas-Congo Cataractes 10 36 South Africa (SA) SA Western Cape Wynberg 1 11 SA Gauteng Pretoria 2 22 SA Gauteng Kempton Park 3 8 SA Gauteng Randburg 4 16 SA Gauteng Wonderboom 5 37 SA Gauteng Germiston 6 13 SA Western Cape Goodwood 7 26 SA Gauteng Alberton 8 21 SA Gauteng Bronkhorstspruit 9 339 SA Gauteng Johannesburg 10 1 Tanzania Tanzania Zanzibar West Magharibi 1 135 Tanzania Morogoro Morogoro Rural 2 136 Tanzania Iringa Iringa Rural 3 16 Tanzania Mwanza Ilemela 4 134 Tanzania Pwani Maﬁa 5 130 Tanzania Zanzibar South and Central Zansibar Central 6 11 Tanzania Arusha Arumeru 7 5 Tanzania Arusha Simanjiro 8 127 Tanzania Kaskazini-Unguja Kaskazini ’B’ 9 8 Tanzania Kilimanjaro Moshi Rural 10 7 Notes: Overall rank is based on counterfactual analysis described in Section 8.1 and the ρs estimated in column (8) of Table 3. Overall key-player (KP) rank is based on the ρs estimated in column (8) of Table 3 as well. The number of districts per country are: Ethiopia 72, Egypt 26 (ADM1 level), DRC 38, SA 354, and Tanzania 136.

A.28 Table P2: Top-Ten Rankings from Policy Experiment 2 (Roads) for Populous Countries

(1) (2) (3) (4) (5) Country Province District Overall Overall Rank KP Rank Ethiopia Ethiopia Tigray Central Tigray 1 17 Ethiopia Tigray Easetern Tigray 2 16 Ethiopia Afar Zone 5 3 70 Ethiopia Tigray Southern Tigray 4 65 Ethiopia Afar Zone 4 4 71 Ethiopia Amhara South Gonder 6 66 Ethiopia Amhara West Gojam 6 9 Ethiopia Amhara North Wollo 8 54 Ethiopia Addis Ababa Zone 5 9 6 Ethiopia SNNP* Konso Special Woreda 10 29 *Southern Nations, Nationalities and Peoples Egypt Egypt Al Qalyubiyah 1 1 Egypt Ad Daqahliyah 1 5 Egypt Asyut 3 19 Egypt Al Minuﬁyah 3 3 Egypt Suhaj 3 18 Egypt Al Fayyum 3 11 Egypt Al Qahirah 3 8 Egypt Al Iskandariyah 3 12 Egypt Al Isma‘iliyah 3 15 Egypt Al Buhayrah 3 7 Egypt Al Gharbiyah 3 2 Egypt Kafr ash Shaykh 3 10 Egypt Dumyat 3 6 Egypt Bani Suwayf 3 9 Democratic Republic of the Congo (DRC) DRC Bas-Congo Bas-Fleuve 1 10 DRC Bas-Congo Boma 2 7 DRC Bas-Congo Matadi 2 2 DRC Kivu Sud-Kivu 4 1 DRC Bas-Congo Cataractes 5 36 DRC Kasa¨ı-Occidental Lulua 6 34 DRC Kivu Nord-Kivu 7 9 DRC Kasa¨ı-Occidental Kasa¨ı 8 21 DRC Equateur´ Equateur´ 9 16 DRC Orientale Ituri 10 15

A.29 South Africa (SA) SA Orange Free State Bloemfontein 1 115 SA Orange Free State Botshabelo 1 27 SA Orange Free State Thaba’Nchu 3 99 SA Orange Free State Dewetsdorp 4 283 SA Mpumalanga Moutse 5 65 SA Gauteng Cullinan 6 190 SA Mpumalanga Moretele 7 352 SA Mpumalanga Mdutjana 7 41 SA Mpumalanga Mbibana 9 114 SA Limpopo Bochum 10 349 SA Limpopo Seshego 10 77 Tanzania Tanzania Kilimanjaro Mwanga 1 116 Tanzania Kilimanjaro Same 2 115 Tanzania Kagera Bukoba Rural 3 22 Tanzania Manyara Simanjiro 4 123 Tanzania Manyara Karatu 5 103 Tanzania Mtwara Masasi 6 72 Tanzania Manyara Mbulu 7 64 Tanzania Mwanza Nyamagana 8 10 Tanzania Mwanza Lake Victoria 8 40 Tanzania Mara Lake Victoria 10 94 Notes: Overall rank is based on counterfactual analysis described in Section 8.2 and the ρs estimated in column (8) of Table 3. Overall key-player (KP) rank is based on the ρs estimated in column (8) of Table 3 as well. The number of districts per country are: Ethiopia 72, Egypt 26 (ADM1 level), DRC 38, SA 354, and Tanzania 136.

A.30